Free DP-100 practice test: 30 questions with answers

The model catalog lets teams discover, compare and deploy foundation models from many providers. Training data storage, billing and notebook hosting are different functions.

Benchmarks plus tests on representative tasks show quality, cost and speed trade-offs. Size, naming and avoiding own-data tests are poor selection methods.

Prompt variants let a team test alternative wording against the same inputs and compare results. Chat history backups, weight backups and hidden prompts are unrelated.

Chunking produces pieces small enough to embed and retrieve precisely, which improves relevance. Printing, zero-size files and per-chunk model training do not describe chunking.

Azure AI Search holds the index of chunks and embeddings and returns relevant context for a query. It does not train the model, host the interface or replace the model.

Fine-tuning needs prompt and desired-response examples in the expected format, and reviewed quality matters. Unstructured text, unreviewed self-outputs and screenshots do not meet that need.

Groundedness checks whether claims are supported by retrieved passages. Word counts, utilisation and popularity do not measure grounding.

Frequently changing content can be re-indexed without retraining, and retrieval supports citation. RAG does not teach style through weights, remove access control, or avoid evaluation.

Jobs reference data assets as inputs, and the platform mounts or downloads them at run time. Copying data into code folders, reading laptops and embedding values in YAML are poor practice.

A component is a reusable step with defined inputs and outputs. Workspace settings, dashboards and model storage are different features.

The signature defines the input and output schema, which callers and deployments rely on. Approvals, endpoint prices and training hardware are not part of the signature.

A blue-green style rollout deploys alongside the current version and shifts traffic gradually, which limits the impact of a fault. Deleting first, switching all traffic at once and disabling authentication each raise risk.

Batch endpoints score large volumes asynchronously, which suits scheduled workloads. Low-latency interactive requests need online endpoints, and the other options do not describe a scoring need.

The failed step's logs show the error and where it occurred, which is the right starting point. Deleting the workspace, hiding failures and resizing storage do not diagnose the failure.

Job logs show script output and environment messages, which support troubleshooting. Logs are retained for review, they do not contain model weights, and they are not hidden from the team.

Testing before traffic shift and authenticating callers are standard safe practices. Full traffic without checks, public model files and permanent zero scale each create a risk or a non-working endpoint.

A compute instance is a managed development workstation for an individual user. Clusters, inference clusters and hand-maintained virtual machines serve other purposes.

A datastore is a connection to a storage service that jobs use to reach data. Model artefacts, logs and notebook copies are stored in other places.

A uri_folder asset references a folder, which suits collections of images or files. uri_file references a single file, mltable describes tables, and environments define the software.

A custom environment is needed when the job requires packages or an image that the curated environments do not provide. Removing compute targets, disabling logs and skipping registration are not reasons to build one.

A registry shares assets such as models, environments and components across workspaces. Logs, account lists and Git credentials are workspace or identity concerns.

Git gives change tracking, review and rollback for scripts. Storing models in Git, removing compute and committing data automatically are not benefits of the integration.

Understanding the dataset's structure and format guides the tools and compute needed. Buying hardware, deploying first and naming the model before looking at data reverse the sensible order.

Regression predicts continuous values from tabular features. Classification predicts discrete labels, segmentation works on images and clustering has no target.

The Responsible AI dashboard helps assess fairness, error patterns and explanations before deployment. The other options describe cost reporting, disaster recovery and labelling tasks it does not cover.

Grid sampling exhaustively tries every combination of discrete values. Random and Bayesian sampling choose values differently, and early sampling is not a sweep method.

The Bandit policy uses a slack factor or slack amount relative to the best run. Median stopping uses medians, while random and manual stopping do not compare against the best run.

MLflow tracking records parameters, metrics and artefacts for each run so that runs can be compared. Keeping only the model, only costs or only names misses that comparison data.

Automated ML is useful for quickly finding baselines and comparing preprocessing and algorithm choices. It does not replace responsible AI review, engineering review or the labelling of data.

Automated ML supports NLP task types such as text classification and named entity recognition. Image classification, forecasting and sensor anomaly detection do not apply to plain text in that way.